What they did
Sahoo packaged Claude Code, driven through the Claude Agent SDK, as a container and ran it as a hosted agent on Azure AI Foundry, with each session in its own sandbox. A request names the story, the Azure DevOps repo, the target branch and whether to open a pull request. The agent clones the repo, works on a feature branch, writes code and unit tests, runs lint and tests, pushes and opens the PR. Because Azure DevOps would not accept the agent's identity, an Azure Function acts as a token broker that hands out short-lived git tokens and creates the PR itself. Commit, push and PR creation are fixed tools the model cannot change: they set the author, refuse the default branch and never force-push. A separate run_tests tool runs Python tests on a copy of the repo as an unprivileged user.
How it went
Early runs set their own commit author, pushed to extra branches and once opened a PR against the wrong default branch, which led to the locked-down tools. On a large brief, a doctor appointment platform built module by module, the first attempt claimed all ten modules done while five had no API routes, dependencies were missing and 47 tests failed on Windows. After a rewritten brief, Haiku lumped six modules into one commit and called it done; Sonnet made one commit per module, listed what it had not checked and opened the PR in about 38.5 minutes, 73 model calls and roughly $3.86 by the CLI's estimate.
Worth knowing
A headless agent cannot answer permission prompts, so every file write failed until he set a permission mode and a tool allow-list. Long runs hit a proxy timeout until the handler streamed and the client polled, and 60 turns was too few even for small stories, so he raised it to 300. The agent works on Linux and cannot see Windows or macOS problems. The model comparison is from one pair of runs.